Automatic train operation requires a control strategy capable of maintaining speed-tracking accuracy, passenger comfort, operational safety, and energy efficiency under continuously changing railway conditions. Conventional control approaches can provide effective regulation, but their performance may be constrained when train dynamics, disturbances, operational constraints, and nonlinear relationships become difficult to represent through fixed control rules. Reinforcement learning provides an alternative paradigm in which an intelligent controller learns operational decisions through interaction with an environment. This study develops a conceptual deep reinforcement learning approach for optimal automatic driving of urban rail trains, with particular emphasis on Deep Q-Network (DQN)-based decision-making. The proposed approach integrates train-state representation, discrete driving-action selection, reward-based optimization, and operational constraints into a unified automatic driving framework. The literature indicates that iterative learning control, self-anti-disturbance control, Q-learning, policy-gradient reinforcement learning, and comfort-evaluation methods provide important foundations for intelligent train control. However, these approaches also reveal a need for a more integrated framework capable of balancing speed tracking, energy consumption, and ride comfort. The proposed methodology therefore formulates automatic driving as a sequential decision-making problem and defines a multi-objective reward mechanism incorporating tracking error, acceleration variation, energy consumption, and operational constraints. The resulting framework provides a theoretical basis for adaptive and optimization-oriented automatic train driving and establishes directions for future experimental validation using real or simulated urban rail operating data.